Pseudo-Labeling and Contextual Curriculum Learning for Online Grasp Learning in Robotic Bin Picking

• DOI: 10.48550/arxiv.2403.02495
• Publication Date: 2024-03-04
• Authors: Huy Viet Le; Philipp Schillinger; Miroslav Gabriel; Alexander Qualmann; Ngo Anh Vien

The prevailing grasp prediction methods predominantly rely on offline learning, overlooking the dynamic learning that occurs during real-time adaptation to novel picking scenarios. These scenarios may involve previously unseen objects, variations in camera perspectives, and bin configurations, among other factors. In this paper, we introduce a approach, SSL-ConvSAC, combines semi-supervised reinforcement for online learning. By treating pixels with reward feedback as labeled data others unlabeled, it efficiently exploits unlabeled enhance addition, address imbalance between by proposing contextual curriculum-based method. We ablate proposed approach real-world evaluation demonstrate promise improving tasks using physical 7-DoF Franka Emika robot arm suction gripper. Video: https://youtu.be/OAro5pg8I9U